Interview with Simon Peyton-Jones
cs.cmu.edu
cs.cmu.edu
Still, I wish I could see more info on this. At what point does the additional cognitive burden of advanced type system features become a worthwhile tradeoff for program correctness? It seems to me that this depends wholly on the complexity of the program.
Further to that point, the most complex programs I can think of (perhaps you may be able to offer other opinions, which I welcome) are AAA game engines. What are the reasons why the big engines out there are not using higher-kinded types, dependent types and the like? Is this just because of pragmatic issues such as the languages the developers learned in school not supporting these features, or because here-to-date functional languages supporting these features lacked the appropriate throughput of C/C++, where one can layout data for cache-efficiency?
Also, it's sometimes possible to use high-flying type trickeries in a local way that is not directly visible to the outside, but highly increase the safety inside a library.
In all cases, Rust is probably the first language which both has a type system rich enough to start doing advanced type trickery and the capacity for fine-grained control over memory layout. I know they have a fairly vibrant gamedev community but I don't know the details. Do they use all the fancy things such as phantom types, complex traits, and all that?
I'm inclined to believe a compiler is better suited for a type system like Haskell's than e.g. a mobile app or a game engine is, but maybe someone with more domain knowledge might correct me here.
GC is also a very blanket term. There are many different types. From what I understand the Unreal GC gives you a lot more explicit control over how and when it runs.
Having written OpenGL applications in Haskell (not games) I can say that it has no such feature. It's GC is also vastly different as immutable functional languages generate a LOT more garbage than normal applications (measured in GB/s).
GC is not the impediment. GC as it is currently implemented in the majority (if not all) advanced type system functional languages is.
However Unity is improving their performance, you should have a look at HPC# presented at GDC 2018.
https://github.com/tweag/ghc/tree/linear-types
However, to allow the compiler to optimise pure operations to impure ones, I don't think linear types are sufficient. You need something like uniqueness types.
x = doSomething(x) + 1
int x2 = doSomething(x) + 1
let x' = doSomething x + 1
As a professional Haskell programmer, I find the cognitive burden to be lower in Haskell than in, say, Java, where I have to do a lot more bookkeeping about design patterns and how they're glued together than in standardish Haskell where compositional forms fall into a compact set of powerful concepts amenable to reasoning.
I say standardish Haskell because the sweet spot in my experience is a few lightweight GHC extensions but mostly shunning some of the seriously experimental stuff in the language. For example, i agree with your doubt when it comes to dependent types and in particular singletons, a halfway house implementation that can be used in Haskell today.
Some of my coworkers had unsuccessfully attempted to write mobile games in Haskell. That was 10 or so years ago and many of the technical hurdles that impeded them then are no longer there.
The only major one I know of at the moment that prevents Haskell from being used in a AAA game engine is a guaranteed low latency garbage collector. I expect someone to implement such a thing for Haskell in the next 10 years.
The space is moving fast, our understanding of how to write big Haskell apps has advanced drastically since when I first started using the language! I expect big things to come in the next couple of years.
This is not to say that the space is getting rewritten all the time, it's just that more useful concepts are being discovered and matured. For example, Applicative only 11 years ago and scalable FRP like 4 years ago or so.
I should say what type concepts compose my compact tool set:
* higher order functions with syntax optimized for using them.
* algebraic data types (sparingly generalized algebraic data types)
* type classes + higher kinds (The synergy is far greater than the individual features)
* monad transformers (the promise of aspect oriented programming actually realized)
>I say standardish Haskell because the sweet spot in my experience is a few lightweight GHC extensions but mostly shunning some of the seriously experimental stuff in the language.
Are you able to say which GHC extensions you use and why?
This extension fixes a defect in the spec as far as I'm concerned:
* ScopedTypeVariables
Syntactic extensions that make certain code forms lighter:
* LambdaCase * MultiWayIf
* OverloadedStrings * RecursiveDo
* TypeApplications * PatternGuards
* KindSignatures
Code Generation:
* DeriveFunctor * DeriveTraversable
* DeriveFoldable * DeriveGeneric
* StandaloneDeriving * GeneralizedNewtypeDeriving
Extensions to the type system. These can be misused so they should be respected.
* MultiParamTypeClasses (Can weaken type inference)
* FunctionalDependencies (Helps type inference with MultiParamTypeClasses, these two together compete with TypeFamilies for functionality)
* FlexibleInstances (Can be a bit sketch, actually) * FlexibleContexts
* UndecidableInstances (Surprisingly mostly benign)
* TypeFamilies (Type level programming should be kept to a minimum. Can kill type inference)
* GADTs (Very powerful, sparingly appropriate)
* RankNTypes (Higher rank types are very expressive but have much worse type inference)
Unpleasant extensions that cope with the realities of serious engineering:
* TemplateHaskell
* CPP
Very Spicy extensions that I avoid unless they're really really appropriate:
* PolyKinds * TypeInType
I am mostly neutral or unfamiliar with the others. The only extension I vehemently oppose is RecordWildCards because it is a binding form that doesn't mention the names it binds. It can get really confusing!
I really appreciate it that you took the time to write that out.
Most of these extensions have been around for like 10 years now and are well understood. They integrate well with the language without impacting Haskell2010 code. For example, RankNTypes or any of the code gen ones.
I shouldn't have said "a few" though! There's lots of light extensions. The only ones I consider heavy are some of the type system ones, CPP and TemplateHaskell. TypeInType is the heaviest by far.
I don't think Java is a meaningful competitor. It's a language completely based on 90's hopes that OOP was a good idea. Turns out it is mostly good to program simple things as complex "ravioli code". Better compare against languages that games, kernels, and compilers are typically implemented in.
> monad transformers (the promise of aspect oriented programming actually realized)
How do monad transformers realize aspect oriented programming? In my experience they lead to pretty verbose code and lots of boilerplate. I think the way to achieve "aspect-orientedness" (I think this name is based on the same insight that lead to names like "separation of concerns", "cohesion", or "cross-cutting concerns") is simply to draw modules boundaries by shape of data, not in an OOP style where most objects do a million different things (etc. cat must eat, walk, sleep, meow...)
I find the dismissive tone rather amusing.
A large majority of the code running on our planet today is OOP.
You could argue that there might be better ideas out there, but OOP is certainly an idea that has not only proven itself to be tremendously useful, but that has also been able to adjust and adapt through decades of changing requirements. It's pretty much the only software paradigm that's survived for that long.
In OOP >50% of the LOC is just stupid bureaucrazy, setting up object graphs in the name of "isolation" (the irony), half-initializing fields, conforming to the right interfaces etc. This is completely meaningless, do-nothing code. Worse, it gives the illusion to remove some contextual dependencies this way, but the code never seems to work outside of the context it was created in. It's only much harder to read because the context is files away.
OOP is the wrong-minded idea that a program should be a bundle of many "self-contained" objects. But that's wrong, we're writing ONE program here, not thousands. It tries to repair this wrong idea with inheritance (which is at least as bad an idea).
And it makes it really hard to cope for "cross-cutting concerns", which are actually 90% of all we care for, not just a side concern. The complexity is in the edges (i.e. how is information moved/transformed), not in the objects!
OOP mostly survived where performance / architectural scalability is not super important (e.g. Python or similar scripting languages, where it enables dynamic typing). And it survived where the big money is, but not necessarily technical competence (where it enables Object-verb type code completion).
That relates to OOP as in languages like Java - not Alan Kay's idea of OOP, which he emphasizes was very different, but I still don't get what's the idea :p
> A large majority of the code running on our planet today is OOP.
Good example code base?
> It's pretty much the only software paradigm that's survived for that long.
Maybe check on your history? Many people are totally happy with procedural programming.
Do Haskell programmers not create mocks to test external components?
> OOP is the wrong-minded idea that a program should be a bundle of many "self-contained" objects. But that's wrong, we're writing ONE program here, not thousands.
The number of programs isn't the relevant metric. Complexity is. Any complex system is going to trend toward modularity. Modularity requires standard interfaces, which inevitably lead to bureaucracy.
A 1MM line Haskell program is going to be similarly bureaucratic. There are going to be standards you have to adhere to in order to play nice with the rest of the system. That's what typeclasses are, after all.
OOP is traditionally defined by three things: polymorphism, encapsulation, and inheritance.
Polymorphism: Modern Non-OOP languages can also be polymorphic, so that's no longer a differentiator.
Encapsulation: You definitely want encapsulation if your data is mutable.
Inheritance: This is the only truly problematic feature, and it's certainly abused, but it has its place. I don't always want to compose and delegate 20 methods when I just want to change the behavior of one.
Haskell had ad-hoc polymorphism way before Java was a twinkle in its creator's eye. Before Haskell, Miranda (the language Haskell was based off of) could have kicked Java's polymorphism to the curb. Neither Java nor OOP invented polymorphism. If anything, they butchered it by introducing subtyping.
> Do Haskell programmers not create mocks to test external components?
The equivalent in Haskell would be having some kind of 'effects' system. An effect system differs from a mock object in that it limits in its totality what kind of interactions can take place. Typically, each layered effect also has a set of laws. Pure interpreters can be written for these effects, but the impure (i.e., real-world) interpreters are not privileged in their consumption of this effect. The pure interpreter also provides a proper implementation, such that you should be able to replace your real program with all pure interpreters, supply all your input at once, and still have a correct program. In other words, a Haskell program is typically polymorphic over which effects it uses in a way that other languages simply aren't.
> Encapsulation: You definitely want encapsulation if your data is mutable.
Again, Haskell, Miranda, and Lisp had encapsulation long before OOP came about, and Lisp has mutable data.
And dynamic dispatch is already part of Smalltalk and Objective-C, I believe.
Picture an object-oriented procedural language like Java, and then make everything immutable; for every 'void' method, return an updated copy of 'this'; for every non-void function, return a tuple of the updated copy and the value you were going to return. No change to const methods, obviously. And you're done, and you can still take advantage of encapsulation, polymorphism, and inheritance without any hoops. You can even do it in Java itself as a style thing without too much effort and only a modest amount of boilerplate. Alternatively, it's not that much work to build this same thing out of lambdas and dictionaries, if your language has those but not objects (adding mutation into that object system would then be trivial) (and of course, you can build dictionaries out of lambdas and lambdas out of objects, if need be).
Do we create test substitutes, alternate implementations of the same interfaces? Yes. But dedicated mocking frameworks are crazy. In Haskell-like languages if you want an implementation of interface foo that returns bar when called with baz, you just... write an implementation of interface foo that returns bar when called with baz. If the easiest way to do that in your language is some kind of magical reflection-based framework, something is very wrong with your language.
Languages have strengths and weaknesses. Certain tasks are easy in some languages, and certain other tasks are not. Throwing ones hands up and saying "something is very wrong with your language" because one is not familiar with a technique or tooling popular in another language is immature, IMO.
For example, if you explain how Debug.Trace works in Haskell to programmers familiar with Java, and they'd call it crazy.
Agreed, but we mustn't fall into the fallacy of assuming that means no language can ever be better or worse than another. There are good and bad language design choices, and "an implementation of interface foo that returns bar when called with baz" is not some obscure specialized feature, it's the basics of general-purpose programming.
> Throwing ones hands up and saying "something is very wrong with your language" because one is not familiar with a technique or tooling popular in another language is immature, IMO.
I'm very familiar with the techniques and tooling of mocking frameworks. I do not make these claims lightly.
Agreed.
> "an implementation of interface foo that returns bar when called with baz" is not some obscure specialized feature
It isn't some obscure specialized feature in Java either.
Foo foo = new Foo() {
public Bar method(Baz baz) {
return new Bar("bar");
}
}
My Haskell is rusty, but given Haskell psuedocode like:
main :: IO ()
main = do
f <- foo
if (f == 1) then
putStrLn "Got 1"
else
putStrLn "Didn't get 1"
foo :: IO Int
-- ...
This is not a snark; I am truly interested to know how Haskell gets rid of the need to bind alternate implementations of an interface for testing purposes.
To the extent that that's true, fine. I'm sure I see a lot of developers writing something along the lines of:
Foo foo = mock(Foo.class)
foo.when(method(any())).return(new Bar("bar"));
> use dynamic bytecode generation (not reflection, BTW)
How is it not reflection ("the ability of a computer program to examine, introspect, and modify its own structure and behavior at runtime")?
> What the makes the language "worse" to require or allow doing this?
Reflection or code generation means stepping outside the language and its usual guarantees - any time the programmer is forced to do it it's because the language didn't provide a good way to solve the problem within the language itself. It means you can no longer e.g. extract common expressions, because they don't necessarily mean the same thing; if you have some common mock setup code you can't just blindly automatically extract method, you have to think carefully about when the mocks get instantiated and when the expectations are set.
> how would you test that the two branches of main behave appropriately?
> This is not a snark; I am truly interested to know how Haskell gets rid of the need to bind alternate implementations of an interface for testing purposes.
It doesn't - as I said, you still write test implementations of your interfaces. What it does remove the need for is mocking frameworks, which people use in e.g. Java either because implementing the interface the normal way in the language is more effort (not a problem in Haskell), or because they want to test the specific interactions with the object (e.g. "verify that method foo was called twice") because those methods are used for side effects.
Haskell avoids that one by making it easier to represent actions as values; you can use e.g. a free monad to represent actions that will be performed later, so rather than testing that your complex business logic method called deleteUser(userId) on your mock, you instead test that it returns a DeleteUser(userId) value. To a certain extent you can do this in Java too ("the command pattern"), but without higher-kinded types you can't have a standard implementation of e.g. composed commands or standard methods for working with them, so it gets too cumbersome to really do in practice.
Even in Java you wouldn't want to use mocks for testing methods that operate on simple datatypes: to test e.g. a regex find method, you wouldn't pass in mock strings, you'd just pass in real strings and confirm that the results were true or false as expected. A language like Haskell just expands the space of what you can test in the same these-inputs-these-outputs way, by making it easier to represent more things as values.
Why is this a bad thing? How is this different from, say, using Template Haskell?
> It means you can no longer e.g. extract common expressions, because they don't necessarily mean the same thing; if you have some common mock setup code you can't just blindly automatically extract method, you have to think carefully about when the mocks get instantiated and when the expectations are set.
I have never found the existence of tests using mocks being a hindrance to refactoring in Java. Can you provide a more specific example?
PS: > How is it not reflection ("the ability of a computer program to examine, introspect, and modify its own structure and behavior at runtime")?
Agree that it is reflection, by that definition.
Because it shows the language could be a lot better. A common, basic task shouldn't be so much easier outside the language (via reflection) than inside it.
> How is this different from, say, using Template Haskell?
It's the same thing. To the extent that people feel the need to use Template Haskell to do basic and common things, something is very wrong with Haskell.
> I have never found the existence of tests using mocks being a hindrance to refactoring in Java. Can you provide a more specific example?
I mean you can't refactor the test itself. Just basic things like: if you do expect(methodOne(matches("someComplexRegex"))) ; expect(methodTwo(matches("someComplexRegex"))), if you try to pull matches("someComplexRegex") out as a variable you'll break your test (you have to make it a function instead). You can't move an expect() above or below another method call without checking to see whether that was the method it was testing. Individually these things are trivial, but they add up to a chilling effect where people don't dare to improve mock-based tests a little as they work on them, so they end up as repetitive code with subtle variations, just like main code would if you never refactored it.
In my experience, I have not come across such effects. People understand the purpose, strengths, weaknesses and limitations of the libraries they use and try not to "cut against the grain".
> people don't dare to improve mock-based tests a little as they work on them, so they end up as repetitive code with subtle variations, just like main code would if you never refactored it.
I understand this is a subjective preference, but I try not to refactor test code too much. I strive to make my test code not have branches ("if-less code" as some people call it). Sometimes this lead to slightly more verbose code, but in the long run I have found it useful for my test code to be rather boring.
----
I now understand the point you are making, and agree with it technically. I don't agree that those technical points lead to the social effect you call out, because I have not come across it.
Overall, Java makes two bad design choices - nullability by default, and mutability by default. But in the codebases I have worked with in the last few years I, and my colleagues, tend to not opt in to these defaults. This leads to pleasant, testable codebases to work with. We also enjoy acceptable performance, good tooling, easy-to-reason memory usage, great library ecosystem etc.
There are a few more, even today: using a weird secondary type system to track what kind of errors can occur (checked exceptions), classes being non-final by default, universal methods (every method in java.lang.Object except possibly getClass() ought to be moved to interfaces that user-defined types have the choice of not implementing), a bunch of syntactic ceremony around blocks (braces required everywhere, "return" being mandatory) which gets even worse once you want to move away from mutability by default, variance at use site only, no sum types, no HKT...
> This leads to pleasant, testable codebases to work with. We also enjoy acceptable performance, good tooling, easy-to-reason memory usage, great library ecosystem etc.
Sure. There are a lot of good things about the Java ecosystem, and if you see the language as a modest, incremental step over C++ then it is an improvement on that front at least. At the same time I do think ML-family languages - even ML itself - offer a lot of advantages especially if we're talking about them just as languages. In practice I work in Scala and gain most of the advantages of the Java ecosystem but with a language that has most of the advantages of Haskell as well.
>Good example code base?
The majority of code running Amazon the retail site and AWS. Large swathes of Microsoft and Google online services. Most code running bank back office and many trading systems. Likely 90%+ of the line of business applications run by the Fortune 500. Most desktop GUI programs. Should I keep going?
If popularity is a valuable metric, then javascript is probably the greatest language ever invented (sigh).
It's bad of me to react to a single poster because I've seen many similar well reasoned arguments about why OOP is bad, but refer to something I would not call OOP -- I would just call it bad programming.
To put it on it's head, there is a lot of terrible code out there. I wouldn't say that code whose authors believe it is OO is measurably worse than any other code I see. What I agree with is that it also isn't any better (which is often a surprise to those authors). I've seen some good OO code. I've seen some good procedural code. I've seen some good functional code. I've seen some good declarative code. But I've mostly seen terrible examples of all of them :-)
One of the very unfortunate things that happened in the late 80's and 90's was the idea we should write self contained components that we would somehow plug and play all over the place. Despite the many, many horrible systems we wrote like that, the idea refuses to die. Some people believe that this is OO. They are wrong :-)
We have exactly the same problem with "unit testing". Some people mistakenly believe that a "unit" is a piece of code taken in isolation. Then they think, "Hey... it's isolated... what else is isolated? Oh yeah! Objects!" So an object becomes a "unit" and it's tested in isolation... How do I isolate it? Oh yeah! I'll make fake objects for it to talk to.
Yeah, it's a tyre fire. But even though it is popular and even though it is popularly called OO, I think it's a bit unfair. It would be like saying, functions are procedures that return values so the only difference between functional programming and procedural programming is that you always have to return a value. Yeah, it's super wrong, but it's so simple that you could convince a whole bunch of people it's true and then complain about how useless FP is.
I realise that you realise that Alan Kay's idea of OOP is different, but the key part is "I still don't get what's the idea". When you find out, it would be nice to find out your reaction. Otherwise it's really just a strawman rant about how terrible mainstream programmers are (and we are.. terrible, that is ;-) ).
And pretty much nobody really understands what exactly is Alan Kay's vision, simply because it's rather vague and somewhat removed from practical reality. He seems to want "extreme late binding" (which need not be bad) and asynchronicity (Actor Model? can remove control and thus be problematic). He also has some vague idea of extreme parallelism that seems just very far removed from computational reality today. I think it's more a philosophical idea of how the physical and biological world could be seen as concurrent processes. I don't want to say his are bad ideas, and he is obviously an extremely intelligent and educated guy, but at the same time he also does not seem like an experienced programmer from a practical standpoint - but more of a visionary. So that's my idea, and if you have a different one, feel free to head over to the C2 wiki to convince yourself that most people don't really know what he means.
> Good example code base?
It would be easier to give examples of non-OOP crucial programs, for they are far less.
OOP codebases: -- Almost all major AAA games.
-- Almost all GUI systems.
-- All major browsers.
-- Almost all video editors (NLE),
-- Almost all audio DAWs
-- All of Adobe's Suite
-- Almost all 3D and CAD tools
-- All office suites (MS Office, OpenOffice, iWork),
-- All major IDEs (Visual Studio, IntelliJ, XCode, Eclipse)
-- Most of Windows and OSX standard libraries
-- Clang (and GCC now that they went to C++? Not sure if they use OOP)
-- the JVM
need we go on?
Nah. As far as I hear, most of them moved away from OOP a long time ago. "Component systems" have been hot for more than five years.
> Almost all GUI systems
Come on, what a GUI does vs a non-GUI "realtime" app is pretty trivial. A bit of layouting, routing a few input events. I'm not saying that e.g. building a scene graph or box model would be the wrong thing (there are other ways as well), but yeah... I certainly don't think that inheritance makes GUIs easier. Interfaces/Function pointers/dynamic dispatch? Yes, you might want that for decoupling a few containers of heterogeneous data/behaviour. But that's hardly a monopoly of "OOP", and also you don't need that in most places.
The other projects, don't know them. Again, I'm not against abstractions per se (the Stream abstraction is one I regular use, although it does require unwrapping for proper error handling. And I have a constant number of "objects" in each of my programs, namely modules with global data). But OOP culture, especially objects-first mentality and fine-grained data encapsulation, gluing implementation to pristine data - I believe it does only harm and leads to plain wrong program structure and overcomplicated code to jump into all these boxes. I prefer gliding through arrays :-)
Nope, that's just hype. Good ole C++ still rules the day, except for specialized (and smaller in scope and speed needs) games.
>Come on, what a GUI does vs a non-GUI "realtime" app is pretty trivial.
You'd be surprised. A guy is not just "call x program with some flags" as you might believe.
Something like a NLE editor for example or even an IDE can have GUI needs that go far beyond hundreds of thousands of LOC...
(Not to mention that I was referring to GUI libraries themselves, complex beasts on their own, not GUI code as used by applications to build their GUIs).
Sure, it's C++, and in my perception much C-style C++. And C++ != OOP.
> (Not to mention that I was referring to GUI libraries themselves, complex beasts on their own, not GUI code as used by applications to build their GUIs).
I will admit I haven't written a NLE program, but integrating complex logic with a complex and hard to understand GUI framework like Qt, meaning there are a lot of states to synchronize, is a lot of inessential complexity. I'm pretty sure it's much easier when you only use GUI primitives and do most of the coding on your own. E.g. a standard box layouting and event bubbling algorithm can't be that much work, and it's MUCH MUCH easier to do it on your own and choose the appropriate structure, instead of writing many helper classes trying to bend the rigid framework.
Taking the example of a NLE program - it has lots of domain-specific state which you must absolutely understand if you want to write such a program. And you absolutely must have a vision how this state should be reflected on the screen. Choosing how to do it is a lot of work. Actually drawing it should be the smaller amount of work, by far. Just separate the state from the GUI library. If you scatter it over thousands of objects, inheriting implementations that you don't own and don't understand - well, of course! that's really hard.
Sorry, meant to write "good old OOP still rules the day".
My point was rather than just the essential GUI functionality and interactions for such a program can be very complex.
In other words, when to show this or that, how to structure code to show it fast, how to present it best, etc is also essential logic (not some afterthought on top of the domain logic), and it can also be super-hard to code (depending on the kind of program).
It's pretty trivial, at least not any sort of GUI-specific hard problem. There is only a low number of objects on any given screen. Don't raster pixel-by-pixel on the CPU, of course.
For anything above that, it can hairy real fast.
Does this include Smalltalk and CLOS codebases?
I think his idea of OOP is something very close to what we call actors these days.
It's a fractal design where it's 'computers' / objects communicating with messages all the way down.
Perhaps the only one that’s been the most popular for that long.
Functional programming predates non-functional programming - Turing's papers and thesis were published (at least) a year after Church's papers on lambda calculus.
Type systems in FP run decades ahead of type systems in regular programming languages. For example, simple type system for FP was published in 1948 and it was (more or less) equivalent to Fortran's type system (1958). The type inference was published in 1968 by Hindley and Milner adapted his algorithm to more "efficient" mutable state in 1978. Type inference come to mainstream languages only in what? 2004?
The algebraic types and pattern matching were born in 1971, the year I born too. These facilities start to appear in mainstream languages no earlier than 2008 if you consider Rust at that time as a mainstream language. And C# acquired them, I think, in 2016 and not earlier.
I boldly and offensively assume that OOP is the only paradigm you decide to care about and thus you consider it "the only software paradigm". I think that it is a very useful position in life, not to care about things you decided not to care about. I do that too.
Just curious... what would you consider a meaningful competitor to Haskell?
Just to lay my own cards on the table: I'd prefer Haskell to almost all other languages if we were only talking about the language (well, the GHC dialect). I do use Haskell quite a bit, but unfortunately I also have to do quite a lot of work in the "enterprise" space where ridiculous things like being able to read e.g. Excel spreadsheets is a big deal. (That is, one cannot rely on people sending those spreadsheets over email to convert to CSV in any meaningful way, so...)
It's often joked that it IS possible to design by committee. The result is Haskell. Otherwise, you get Java.
Games... So, C# then? :)
I know they have some different features and can feel a little different, but they’re much closer to each other than ML and Haskell, say.
I hardly see any overlap between these two concepts.
AOP is about cross cutting concerns and being able easily insert similar functionalities in unrelated areas of the code.
Monad transformers are the hack you need to use to make monads compose.
Monad transformers allow one to slice up the functionality and concerns of a program in a dimension different from slicing of a program into components that are then composed together; e.g. a bunch of functions that call each other or a bunch of objects that pass messages between themselves.
Each monad transformer in a transformer stack adds functionality and concerns that the other parts of the stack don't need to care about. Components can then be written polymorphically in the concerns they care about allowing them to be instantiated wherever the capabilities they need are present. Ultimately a program is instantiated to a particular transformer stack and then it is supplied effect handlers that reduce it to the base monad (often IO). The ability to modify functionality in a cross-cutting way is concentrated in the effect handler which is at the discretion of the call site, not the implementation site.
This is not quite AOP but it's in the same space. It's hard to see in small apps but quickly emerges as the scale of a Haskell application grows.
i'm not sure where you got the idea that they 'make monads compose'. Monads don't compose. It's nonsensical to talk about monad composition in general, because there is nothing about the structure of a monad that allows it to compose (quite the opposite, really). Some things which may form some monads do compose with certain other things that form monads, but this is like saying 'Abelian groups' are the 'hack' you need to make groups commutative -- it's a meaningless statement.
I am not the OP, and not a Haskell programer, but my understanding is that Monad Transformers exist because Monads don't compose, and when you want monadic effects to compose, you use Monad Transformers (the other option is a go crazy writing every combination and ordering of effects you want by hand).
I got this impression from the papers I have read, particularly this one [1]
Monad transformers offer an additional benefit to monadic
programming: by providing a library of different monads and
types and functions for combining these monads, it is possible
to create custom monads simply by composing the necessary monad
transformers. For example, if you need a monad with state and
error handling, just take the StateT and ErrorT monad transformers
and combine them.
[1]: Monad Transformers Step by Step https://page.mi.fu-berlin.de/scravy/realworldhaskell/materia...
I see what you mean.
You are right - the concept of Monad Transformers exist independent of the fact that monads don't compose.
What I meant was that MTs exist in Haskell programs because Monads don't compose. Of course, there probably exists a Haskell program where this is not the case, but I am certain MTs are largely used in Haskell because Monads don't compose.
BTW, the grandparent is not the first to coin "MTs are .. use(d) to make monads compose" usage. The late Paul Hudak et al writes in [1] that:
A ground-breaking attempt to better solve the overall
problem began with Moggi’s proposal to use monads to
structure denotational semantics. Wadler popularized
Moggi’s ideas in the functional programming community
by showing that many type constructors (such as List) were
monads and how monads could be used in a variety of
settings, many with an “imperative” feel (such as in Peyton
Jones & Wadler). Wadler’s interpreter design, however,
treats the interpreter monad as a monolithic structure which
has to be reconstructed every time a new feature is added.
More recently, Steele proposed pseudomonads as a way
to compose monads and thus build up an interpreter from
smaller parts, but he failed to properly incorporate important
features such as an environment and store, and struggled
with restrictions in the Haskell type system when trying
to implement his ideas. In fact, pseudomonads are really
just a special kind of monad transformer, first suggested by
Moggi as a potential way to leave a “hole” in a monad
for further extension.
[1]: Monad Transformers and Modular Interpreters (http://haskell.cs.yale.edu/wp-content/uploads/2011/02/POPL96...)
Shockingly, our understanding of monads and monad transformers has advanced in the last 23 years.
@danharaj, of course!
At least enough that when someone says "MTs exist to compose monads", to understand it as "MTs are used in Haskell programs to compose monadic effects".
I hate the way monads in general are talked about in the Haskell community (I also do not like the language for other algebraic terms). In particular, it is common to say 'data type X is a monad'. This statement is completely nonsensical. A monad is a thing along with some operations. Haskell data is a concrete description of how to store a thing. By itself, a piece of data has no operations associated with it. There is no way a particular data type could be a monad, since a data type is just a thing.
The proper terminology is 'X forms a monad along with these functions' or 'X has an instance for the Monad type class' (this is different, because the monad type class allows for a subset of monads, not general monads). I am on a personal mission to rid the Haskell world of sloppy language because it confuses beginners, in my opinion. For a long time, I thought monads were a thing. Then I realized they're just an algebraic structure, like groups or rings.
When we say things like 'monads don't compose' it makes it seem like there is some deficiency in the idea of a monad that makes them not compose, rather than a realization that monad composition results in zero or more ways to combine two arbitrary monads. Thinking just the former makes you wonder if monads ought to be fixed. Realizing the latter means you realize that the lack of a fix indicates that putting two monads together can accomplish a variety of tasks, only one of which is likely suited to your use case.
I think the reason for this is that in Haskell we model algebraic structures using typeclasses, which are dispatched by type. So we say “X is a monoid, with mempty and mappend defined as follows”:
instance Monoid X where
mempty = …
mappend = …
instance Monoid X emptyX appendX
emptyX = …
appendX = …
Yes, monad transformers exist to let you do that. They're the thing that makes it possible to insert an effect in a code area that's using other, unrelated effects.
Cross cutting concerns allow you to insert a certain line of code in "file 1 at line 223 and file 2 at line 400" because these lines match a certain type safe regexp.
Monad transformers accomplish nothing remotely close to that.
> Cross cutting concerns allow you to insert a certain line of code in "file 1 at line 223 and file 2 at line 400" because these lines match a certain type safe regexp.
What you're describing is the implementation details of how AOP works. Cross-cutting concerns refers to the problem statement: secondary concerns that need to be addressed in the same way in different parts of the codebase with minimal disruption to the primary logical code in those different parts of the codebase.
Also, I've found the cognitive burden to be significantly less using my beam library than writing SQL. The library handles everything. We freely join against queries (which may themselves be the results of joins, aggregates, window functions, etc). The SQL produced is what you'd expect. If you try to do something that is not straightforward on the backend, the library complains at compile time, and can even offer suggestions. We don't even need to worry about NULL, because the library makes heavy use of type families and such to guarantee you won't see a SQL NULL unless the type says so. If you use standard beam operators, we handle NULL sensibly. You can also drop to SQL NULL, if you want to have tighter control, but this causes the types to change, which means you can't do anything silly (like inadvertently throw out rows because your join condition now evaluates to NULL when you didn't want it to). Ultimately, using the type system is wonderful. Best of all, the types are mostly inferred. I rarely write out explicit type signatures. Despite the types carrying significant computation, all of that is given to us basically for free.
Once we have linear types in GHC, the migrations system will be able to check that a migration script is valid from beginning to end before the migration is even run! There's no way you could do that with any current tool I know of.
My guess for why big game engines aren't using higher-kinded types is that they are not available in any common systems level language. The only systems language I can think of with HKTs is ATS, which is a bit obtuse.
W.r.t cache efficiency, I've been working on a library to exploit the Haskell recursion-schemes package to transparently store and read data in a cache-intelligent way. It's certainly possible to do, it's just not necessary for most of the kinds of programs Haskellers write. At my last workplace we used Haskell for protocol parsing, and we were able to parse gigabits per second using just the standard Haskell containers and such. Haskell is quite a fast language, if you use it correctly. You rarely need heavy optimizations. Although obviously for game engines, you probably would.
I find this confusing. You still have all the cognitive burden to figure out what SQL you want to write, but then you have to translate that SQL into Haskell (or any other ORM).
I can see that this would be a benefit in the long run, but it wouldn't be familiar to most devs in the beginning. Similar to learning Haskell, once you get over the hurdle of the concepts then there's less to think about over all but there is still that initial gap.
Certainly the declaritive nature of SQL should be a very good fit for Haskell.
The types go a bit crazier than lists internally, but the aim is minimize the times manual intervention is required
Edit I don't mean this as an insult. I see it with a friend of mine who struggles with problem solving when not mapping to cat types
Yes, I remember a Dan Luu and another developer talking about something similar. The concepts like currying, composition and monads all work so smoothly in Haskell that you want to try to get them to work in other languages. It can make your code in those other languages worse because you're trying to get the language to do something it's not designed to do.
Further to that though then no one understands what you've written.
This is what Dan Luu said [0]:
> One was that I really got into functional programming and used a functional style everywhere I could. Immutability, higher-order X for any possible value of X, etc. The result was code that I could write and modify quickly that was incomprehensible to anyone but a couple of coworkers who were also into functional programming.
Beam is based on the relational algebra, which is the basis of SQL. Of course, it doesn’t enable a monkey to write database queries, but it does enable engineers who understand relational algebra to do so.
If your complaint is that you have to understand relational algebra to use a relational database, then your complaint is well out of the scope of beams domain.
Also, I have no idea what you mean about your friend. We can play the personal anecdote game all day long. When I was a JavaScript developer, I saw my fellow devs constantly testing for invariants that would be trivial with Haskell. We use JavaScript at work now too and they access our database indecently. Their queries do not enjoy the advantages beam provides and we’ll be switching to using a Haskell backend service soon instead
When you have a notation well suited for using these advanced features, all of a sudden they become tractable to reason about. The languages in which AAA game engines are written have notations and semantics that are not well suited to these abstractions; and people continue to use them because they have reputations for efficiency and control that functional languages historically do not. But that’s changing with the advent of languages like Rust and ATS that bring more of the power of functional programming and advanced type system features to bear on low-level efficient code without the need for a GC. (I’m also working on such a language.)
What I’m getting at is that there are some things I do all the time in Haskell because it’s easy, which I don’t do in other languages becuase it’s hard. I know in principle how to get the same guarantees in other languages, but the expressiveness isn’t there: the encoding in the syntax & semantics of those languages is so unwieldy that I rarely bother.
Take for example something simple like algebraic data types and pattern matching. You can encode sum types with only product types (tuples/records) and functions using Boehm-Berarducci encoding, which in OOP languages is called the visitor pattern. But it takes so much more code that I often design things entirely differently to avoid it.
In Haskell I use higher-kinded polymorphism all the time—abstracting over type constructors. That’s possible to encode in C++ with template template parameters, but it’s incredibly unwieldy, produces awful error messages when you do it wrong, and can be invasive depending on what features you need to support. I use GADTs, existential types, and higher-rank polymorphism all the time, for example, passing generic functions as arguments to other functions. That’s possible to encode in OOP languages using interfaces with generic methods, but it can’t be done inline, requiring a couple of helper types and moving the “meat” of an implementation away from the actual method being defined; in Haskell I just add a “forall” with the scope I want.
So I don’t think it depends on the complexity of the program so much as the complexity of the encoding in your language of choice. It’s eminently possible to take advantage of these features in C++, Java, C#, &c. to enforce better program correctness and even gain better performance, but the “wizardry” required is not accessible to the majority of users. Whereas in Haskell, sooner or later just about everyone gains some experience with type-level programming that would be possible but prohibitively difficult to do in other languages, because Haskell was designed to support those abstractions—whether through built-in concepts, extensions, or clever combinations of orthogonal and expressive language features.
The more advanced your programming tools are, the more complex your code can get before you hit your limit and need to refactor. It’s great up til that point, but once you cross the line you are in mind melting territory.
I code in a tiny subset of JavaScript, wrap at 40 character width, without allowing any build tools, and only single-file repos. This probably seems insane to most, but it has a powerful effect: Complexity becomes painful very fast, and so I am forced to refactor aggressively, which causes me to put more effort into good separation of concerns earlier in my implementation process.
It’s not for everyone, but if you enjoy the discovery of medium-sized single purpose modules, I would encourage you to try this, in your language of choice. Just pick a small set of native primitives and stick to them, and libraries written in (close to) the same subset.
As a side note: Some people might be thinking, “huh? Rich type systems and control structures HELP you write single purpose code with clear boundaries.” To which I agree. But I said “medium sized”. Rich toolsets will push you towards a combination of very small and very large API surfaces: small modules that do one arcane abstract thing that has no immmediate value, and then a huge surface which is the entire set of all of those small modules.
Modular subset programming makes both small and large modules awkward... small modules are awkward because you need to include them over and over. Large modules are awkward. It squeezes you from both sides into finding concerns that are good brain-sized nuggets.
I like this idea! I've sort of gravitated toward it myself -- inspired by the exceedingly clear pieces of code that can be found in older books on programming -- but I have yet to make it an expölicit purpose of mine. I may want to try that!
On this project in particular there are tens of thousands of LOC, about 50 csproj files, unit tests, our own regex implementation a plugin framework and dll dependencies from half a dozen other projects. But when you step back from all that and look at what the project really does it's just taking various files and stuffing them into a database. It really could have been done with about 30 isolated 20 line bash scripts.
The biggest enemy in our line of work is the complexity we create for ourselves.
I have similar thoughts as those expressed by the sibling post by danharaj (ohai Dan!), but I'd like to stress a particular point: the type system lowers the cognitive burden, rather than increase it.
By way of analogy, think of it this way: spoken languages are quite complex; you have all sorts of stuff to get right:
* conjugation
* noun genders
* tenses
* participles
* prepositions
* etc, etc, etc...
And there are all sorts of fun, intricate grammatical rules that govern which words are supposed to go where. That's a ton of work!
Now, one could argue: wouldn't it be much easier if we just, you know, made the language simpler?
The problem is that each of these things serves some purpose (well, mostly, anyway -- dunno how I feel about, say, gendered nouns); for example, if we stripped away the ability to convey tense, the language would become simpler, but it would lose out on some important expressive power.
While we're at it, we could also shorten the vocabulary. Picking an arbitrary number, let's say we keep only the top 100 words. That'll probably capture most of the essentials, if what is "essential" is "that which ensures you get your basic needs for survival".
But let's say "love" wasn't in that top 100: how would you tell someone that you love them?
That's no problem: whenever we want to tell someone we love them, we can unpack the definition:
love: a gentle feeling of fondness or liking
Uh-oh -- "gentle" and "fondness" didn't make the cut... I suppose we'll unpack those definitions as well. So now we have:
love: a (mild in temperament or behavior; kind or tender) feeling of (affection or liking for someone or something) or liking
Oh hell, temperament certainly isn't in the top 100, so here we go again:
love: a (mild in (a person's or animal's nature, especially as it permanently affects their behavior) or behavior; kind or tender) feeling of (affection or liking for someone or something) or liking
So ...
... yeah, that got out of hand pretty quickly, and we still have a ways to go.
The next time you tell someone you love them, think about how incredible it is that you have a word at your disposal that immediately conveys something so nuanced. Think of all the meaning that's cram-packed into that one word. And it's only one syllable!
So how does this exercise relate back to more advanced type systems?
A sophisticated type system (like Haskell's) allows me think (and express) thoughts that I otherwise wouldn't be able to. I mean, with inordinate amounts of effort (kind of like the effort I would need to keep the whole, fully unpacked definition of "love" in my head all at once), I might be able to do everything I was able to do before -- it would just make things more difficult, and it wouldn't be practical.
Programming in weaker languages (e.g. Java, Go) often feels like trying to tell someone something as seemingly simple as "I love you", but all I have at my disposal is the ability to grunt and flail my arms around -- that's a simpler system, but man is it a lot of work.
Steele played a major role in shaping Java, and his perspective is that from a Lisp vantage point, so, naturally, he argues for the opppsite of you.
An important nuance is that where a language like Go or Java limits you to only sentences using nothing but the most common 100 words, Lisp makes it really easy to define other words and integrate them in your speech as though they always existed.
Of course, that ability may lead to a lack of canonical definitions and everyone may soon speak their local variant of the language, which can be both a good and a bad thing depending on your priorities.
I think most game programmers have a typical "incremental" mentality, where they need to see progress all the way, even at the learning the language stage. And then studios want easily replaceable programmers.
So by this point you'd probably be left with few ideal languages: Java or C#.
Add the constraint of low level optimizations for performance combined with "zero cost abstractions" so that in theory you can keeps a decent amount of productivity while working with heavily optimized code.
So you're left at... C++
Nothing else gets even close to fitting the requirements.
I imagine that if Rust gets popular enough to satisfy the "easily replaceable programmers" imaginary requirement (because I imagine it never works), you'll start seeing game engins written in Rust.
To be honest, I think Rust will never get very popular because it will tend to have a huuuge gap between "library creators" (that will use the full feature set of the language, understand all the lifetimes tricks, never use garbage collection etc.) and the "library consumers", so people will keep being afraid of popping off the hood. I'd have more hopes with someone creating some kind of "superset of Go" as a next systems language, that would add something like generics, local type inference, and a way to turn off GC. The lowest common denominator always wins, "worse is better" and all that...
Compared to Haskell it has fewer features and is much simpler. Think of it as kind of a subset, but with nicer record syntax and error messages, and tightly coupled to the web.
The upside is you can learn it quick and understand other people's code easily. This makes it more fun to use and feel more productive. There is also a different culture - in Haskell things like template Haskell and Lenses are fairly widely used, but although you can in theory do some Lens stuff in Elm, no one actually does. To give you an idea, as a beginner your introduction to Lenses is this beast http://i.imgur.com/ALlbPRa.png. So that's probably an entire weekend to get your head around. (Big respect to Edward Knett though for creating this and dozens of other libraries) You can get your head around all of Elm in that time.
The downside is that you end up with more boilerplate than Haskell. For example there is no monad typeclass, so you have multiple definitions of "andThen" whereas Haskell you can always reach for the generic >>=
I'm working on a side project and having a lot of fun in Elm
I also strictly required that arithmetic and most other operations require length equality. E.g., you should not be able to compare for equality single bit and bit vector of length 10, as it is possible in C and Verilog. This thing alone proved very effective in reducing errors in the code.
The operations that cross clock domains are also prohibited, etc.
The problem with hardware is that it is really hard to change once you "compiled" it. So you have to plan to get rid of errors early, without too much testing.
The problem with AAA games is their budget - you can throw many people at finding bugs by testing (playing and replaying). So you do not need a tool that allows you to get rid of as many bugs as possible as early as possible. The process employed in AAA games development ensures that there will be not many bugs in the end. It is just that process is different than, say, one in hardware development.
I think that's a false equivalence, because there's only a burden if you're doing something complex. Indeed short scripts can see a lot of benefit from using an advanced type system, since they tend to have interacting effects that can easily lead to misbehaviour if uncontrolled.
> Further to that point, the most complex programs I can think of (perhaps you may be able to offer other opinions, which I welcome) are AAA game engines. What are the reasons why the big engines out there are not using higher-kinded types, dependent types and the like?
My experience of the games industry is that it's full of young developers with a macho performance culture (I appreciate that this doesn't match everyone's experiences). "We're adopting a tool that will help us make fewer mistakes" is a tough sell in any segment of the software industry - it requires a mature culture to get beyond the "just write better code" response - but particularly so in games.
SimulaVR is a 3D window manager (under active development) for VR Linux Desktop.
Interesting, never released Tom Clarke was childhood friends with Simon Peyton-Jones. He mentioned on a few occasions that we should all go learn Haskell.
Sometimes I think, "It'd be cool to define some custom types for Postgres that are Maybe<Text>, Maybe<Integer>, etc." But then I think, "Actually, that's how the built-in types work already!"
...oh you must have replied right as I was deleting.
That's pretty cool, I'm used to things being like the archaic TechDirt commenting system...